Sapere Aude - Dare to be wise

The industry is abuzz with conversations about artificial intelligence (AI) and how it’s going to change everything. AI advocates are of the belief that it will eliminate a lot of jobs because it will be more capable than most people for most work, or at least most so-called knowledge work. Advocates will tell you that AI will enable people to enjoy their lives more because they don’t have to work as much. It is true that we’ve seen significant advances in these large language models (LLMs) and the tooling surrounding them over the past 3 years or so.

AI skeptics, on the other hand, raise concerns about the environmental impacts of these models and the data centers needed to continue this rapid growth. Concerns about the massive scale of copyright infringement that was necessary to train these models. They raise concerns about the quality of the output of the models. Concerns about the impact that other people’s LLM usage is having on their own lives, both at work and at home. AI skeptics don’t particularly seem fearful for their jobs, because the LLMs do not appear adequate enough to compare to skilled humans on many tasks. But they are fearful of society’s obsession with LLMs and the impact that obsession will have on society.

I’ve used LLMs to some reasonable degree of success. I find them quite useful when I want to rapidly prototype an idea or automate some research, e.g. asking a question in plain English and insisting on source citations that I can read has been quite a helpful way for me to identify not just solutions to problems, but new sources of information that I wasn’t previous aware of. But I’ve also seen frontier models produce just absolute garbage with an astounding level of unearned confidence.

If we look at LLMs as a capitalist, the upsides of LLMs feel apparent - computers can work without sleep, without meal breaks, without bathroom breaks, without smoke breaks (except when the magic smoke releases, I suppose), and can produce seemingly accurate output at superhuman speeds. If output is what matters, then why would we pay humans to produce slower output when we can dynamically scale our token spend up and down as needed instead?

Of course, on the surface this neglects quality in favor of quantity and speed. But AI advocates will tell you that if you just chain a few models together and use subagents with different “personalities” (read: markdown files) that check the work of each other, then you can get quality quantity quickly. Just throw more tokens at the problem. To some extent, this works. Of course it increases the amount you spend on the output, but even with subagents you can probably still come out ahead against the cost of humans.

But you probably still need some humans. People who can guide the LLMs appropriately. People who can talk to customers and persuade them to buy your outputs. Someone to make sure you’re not promising your customers cars for a dollar. Since output is what matters, and AI adoption is what drives faster output, you want to select for employees who are making the most use of LLMs. So the LLM leaderboards are born, where we glorify the amount of money spent on using LLMs. Nevermind that many of these humans are simply regurgitating slop with the idea that it is providing some new value and passing the cost of understanding on to whoever has to consume the sloppypasta. This phenomenon is making the rounds at the time of writing under the newly coined term “meat proxy.”

The humans you’ve got left who aren’t just meat proxies are convinced that taste will save them. That knowing the difference between “good” and “good enough” is what will distinguish them from other LLM users. The value is still in the output, but because I have taste, my output will be better. And to some degree, taste will be a differentiator as LLMs reduce the barrier to entry and produce a regression to the mean, one way or another. Fewer people will have developed the capacity to know that difference between “good” and “good enough” and the sheer volume of LLM-generated output will skew everything to a perfectly statistical average. An average that you can feel whenever you visit every vibe-coded website, even if you can’t quite explain the feeling.

The underlying theme of the benefit of LLMs as they pertain to running a business is that what matters is output. The value is entirely placed on the outputs, and any and all value from the process is discarded. It has never been easier, or more dangerous, to delegate understanding. The value of understanding has largely been discarded in favor of the hype of fast and easy outputs that feel correct.

I urge you to not delegate your understanding, though. Sapere Aude. Have the courage to use your own understanding. LLMs can be a useful tool, but don’t give in to cognitive surrender. Admit when you don’t understand something and seek help. Don’t let LLMs take away your ability to learn. If you yield yourself completely to the LLMs, atrophy will creep in and you will no longer be able to do the things you were once great at. Things you once enjoyed will be boiled down to just the value of their outputs, and you will find yourself wondering how to recapture the joy you once felt.

It’s not always about output.

Sometimes, it’s about what you’re capable of.